Code-excited linear predictive coding with low delay for speech or audio signals

Abstract

A code-excited linear-predictive (CELP) coder for speech or audio transmission at compressed (e.g., 16 kb/s) data rates is adapted for low-delay (e.g., less than five ms. per vector) coding by performing spectral analysis of at least a portion of a previous frame of simulated decoded speech to determine a synthesis filter of a much higher order than conventionally used for decoding synthesis and then transmitting only the index for the vector which produces the lowest internal error signal. Modified perceptual weighting parameters and a novel use of postfiltering greatly improve tandeming of a number of encodings and decodings while retaining high quality reproduction.

In the claims: 
     
       1. A method of encoding comprising: (a) receiving a set of input audio samples representative of an audio signal, the set of input audio samples comprising a first portion and a second portion;   (b) applying a first hybrid window to the second portion of the set of input audio samples to generate a first windowed second portion;   (c) generating a set of quantized audio samples approximating the set of input audio samples, the set of quantized audio samples comprising a first portion and a second portion;   (d) applying a second hybrid window to the second portion of the set of quantized audio samples to generate a second windowed second portion;   (e) generating a modified digital signal obtained from a set of gain scaled excitation samples, the modified digital signal comprising a first portion and a second portion;   (f) applying a third hybrid window to the second portion of the modified digital signal to generate a third windowed second portion; the first hybrid window, the second hybrid window and the third hybrid window being represented by w m  (n) according to the equations:   w.sub.m (n)=f.sub.m (n)=bα.sup.-[n-(m-N-1)]        if n≦m-N-1   w.sub.m (n)=g.sub.m (n)=-sin [c(n-m)]        if m-N≦n≦m-1   w.sub.m (n)=0        if n≧m    and wherein N is equal to about 30 and α is equal to about 0.98282 for the first hybrid window, N is equal to about 35 and α is equal to about 0.99283 for the second hybrid window, and N is equal to about 20 and α is equal to about 0.96468 for the third hybrid window;   (g) calculating a first plurality of coefficients from the first windowed second portion;   (h) calculating a second plurality of coefficients from the second windowed second portion;   (i) calculating a third plurality of coefficients from the third windowed second portion;   (j) deriving a first set of predictor coefficients, a second set of predictor coefficients, and a third set of predictor coefficients from the first plurality of coefficients, the second plurality of coefficients, and the third plurality of coefficients, respectively;   (l) outputting the index.   
     
     
       2. The method of claim 1 wherein the first portion and the second portion of the set of input audio samples are mutually exclusive. 
     
     
       3. The method of claim 1 wherein b is about 0.960 and c is about 0.060 for the first hybrid window, b is about 0.989 and c is about 0.048 for the second hybrid window, and b is about 0.932 and c is about 0.092 for the third hybrid window. 
     
     
       4. A method of decoding comprising: (a) receiving an index associated with an excitation vector, the excitation vector being representative of a set of audio samples;   (b) choosing a set of previously quantized audio samples;   (c) applying a first hybrid window to the set of previously quantized audio samples to generate a first windowed portion;   (d) determining a modified digital signal obtained from a previous set of gain scaled excitation samples;   (e) applying a second hybrid window to the modified digital signal to generate a second windowed portion; the first hybrid window and the second hybrid window being represented by w m  (n) according to the equations:   w.sub.m (n)=f.sub.m (n)=bα.sup.-[n-(m-N-1)]        if n≦m-N-1   w.sub.m (n)=g.sub.m (n)=-sin [c(n-m)]        if m-N≦n≦m-1   w.sub.m (n)=0        if n≧m    and wherein N is equal to about 35 and α is equal to about 0.99283 for the first hybrid window and N is equal to about 20 and α is equal to about 0.96468 for the second hybrid window;   (g) calculating a first plurality of coefficients from the first windowed portion;   (h) calculating a second plurality of coefficients from the second windowed portion;   (i) deriving a first set of predictor coefficients and a second set of predictor coefficients from the first plurality of coefficients and the second plurality of coefficients, respectively;   (j) generating an audio signal by gain adjusting and filtering the excitation vector, the filtering being based upon the first set of predictor coefficients and the gain adjusting being based upon the second set of predictor coefficients; and   (k) outputting a signal representative of the audio signal.   
     
     
       5. The method of claim 4 further comprising the steps of: (a) postfiltering the signal representative of the audio signal to generate a postfiltered signal; and   (b) converting the postfiltered signal to a PCM output format.   
     
     
       6. The method of claim 4 wherein b is about 0.989 and c is about 0.048 for the first hybrid window and b is about 0.932 and c is about 0.092 for the second hybrid window. 
     
     
       7. A method for processing an audio signal comprising: (a) receiving a set of input audio samples representative of an audio signal, the set of input audio samples comprising a first portion and a second portion;   (b) applying a hybrid window to the second portion of the set of input audio samples to generate a windowed second portion, the hybrid window being represented by w m  (n) according to the equations:   w.sub.m (n)=f.sub.m (n)=bα-[n-(m-N-1)]        if n≦m-N-1   w.sub.m (n)=g.sub.m (n)=-sin [c(n-m)]        if m-N≦n≦m-1   w.sub.m (n)=0        if n≧m}    and wherein N is equal to about 30 and α is equal to about 0.98282;   (c) calculating a plurality of coefficients from the windowed second portion;   (d) deriving a set of predictor coefficients from the plurality of coefficients;   (e) choosing, from an excitation codebook, an excitation vector based upon the set of predictor coefficients, the excitation vector having an index associated therewith and being representative of the first portion of the set of input audio samples; and   (f) outputting the index.   
     
     
       8. The method of claim 7 wherein b is about 0.960 and c is about 0.060 for the hybrid window.